Google's Gemini 3.1 Flash TTS adds natural-language audio tags for speech control
Google DeepMind
- Models
- Developer tools
- #gemini 3.1 flash tts
- #text-to-speech
- #audio tags
Google announced Gemini 3.1 Flash TTS, a text-to-speech model it describes as its most natural and expressive to date, now in preview via the Gemini API and Google AI Studio, on Vertex AI for enterprises, and in Google Vids for Workspace users. The model adds audio tags — inline natural-language commands that steer vocal style, pace and delivery — plus scene direction, speaker profiles and export of settings as API code. Google reports an Elo score of 1,211 on the Artificial Analysis TTS leaderboard and SynthID watermarking on all output.
What Google announced
Google introduced Gemini 3.1 Flash TTS, which it calls its most natural and expressive speech model to date. The headline mechanism is audio tags: natural-language commands embedded directly in the text input that steer vocal style, pace and delivery, including inline tags that can shift expression mid-sentence. Google says Google AI Studio adds scene direction, per-speaker Audio Profiles with Director's Notes for pace, tone and accent, and export of settings as Gemini API code so voices stay consistent across projects. The model supports 70+ languages and, per Google, reached an Elo of 1,211 on the Artificial Analysis TTS leaderboard and sits in its "most attractive quadrant"; that benchmark and description are Google's citations, not independent tests. Google says all output carries a SynthID watermark, and access is in preview via the Gemini API, Google AI Studio, Vertex AI and Google Vids.
Scenario for a localization lead
A localization team producing narrated product tutorials could try authoring a short script in Google AI Studio with audio tags and speaker profiles, then exporting the settings as Gemini API code to regenerate the same voice across languages. Because the source claims 70+ language support and mid-sentence expression changes, the team could compare whether pacing and emphasis instructions survive translation, rather than judging only a single English clip. Any pilot should stay in preview only if the team's own policy permits preview services for its content.
A small, measurable trial
A developer could run a controlled comparison: take a fixed set of scripts, generate takes with and without audio tags and with longer scene-direction prompts, and record a baseline. Success could be judged by blind reviewer ratings of naturalness and how well delivery matched the intended pace and emotion, plus the share of takes needing manual re-recording. Check failures explicitly — drift in tone across turns, tags ignored or over-applied in non-English output, and inconsistent voice after export. Keep a human editor approving final audio, and verify SynthID detection where provenance matters.
Trade-offs and open questions
Natural-language control lowers the barrier to directing speech but adds prompt-design work: tags that help in one scene may misfire elsewhere, so teams should expect iteration and versioned prompts. SynthID watermarking aids provenance but does not replace consent and disclosure processes for voice use. The supplied text does not state pricing, latency, quotas, licensing or enterprise data terms, so these need checking before committing to production. Teams with strict brand-voice or regulated-audio requirements, or no tolerance for preview-era changes, should not rely on it yet.
Context and limitations
- Based on a short source excerpt and Google's own announcement; the full article and model card were not analyzed.
- The 1,211 Elo figure is Google's citation of the Artificial Analysis leaderboard; no independent testing was reviewed.
- Preview access, pricing, rate limits and commercial terms are not specified in the supplied text.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.